Codex vs Claude Code: which coding agent should your business use?

Codex (OpenAI) and Claude Code (Anthropic) are coding agents that read a codebase, edit files and run commands, and their feature sets are now closely matched. Codex is included in every ChatGPT plan, including Free; Claude Code needs a paid Claude plan or an API account. Choose on your existing subscriptions, data requirements and a trial on your own code.

Most development teams have now tried AI coding tools, and the question has moved from whether to use an agent to which one to standardise on. In the Stack Overflow Developer Survey 2025, 84% of respondents said they were using or planning to use AI tools in their development process, and 51% of professional developers said they used them daily. Yet the same survey found that "AI agents are not yet mainstream": 52% of developers either did not use agents or stuck to simpler AI tools.

OpenAI's Codex and Anthropic's Claude Code are the two agents most UK teams shortlist. This comparison uses each vendor's official documentation and pricing pages, checked on 16 November 2026, plus the Terminal-Bench leaderboard and the Stack Overflow survey. Both products change almost weekly, so confirm any detail that drives your decision.

  • Both tools read your codebase, edit files and run commands, in the terminal, editors, desktop apps and the cloud.
  • Codex is included in every ChatGPT plan, including Free; Claude Code needs Pro, Max, Ultra, Team, Enterprise or an Anthropic Console account.
  • Feature gaps have narrowed: both now offer skills, hooks, subagents, plugins, MCP servers and admin-enforced settings.
  • On the Terminal-Bench 4.1 leaderboard, the top Claude Code entry scored 65.8% and the top Codex entry 60.9%, each within about 3 points either way.
  • For UK GDPR, check where prompts are processed and stored, which plan's data terms apply, and whether you need a cloud route such as Amazon Bedrock.
Key takeaways

What are Codex and Claude Code?

Both are agentic coding tools: you describe a task and the agent reads the code, edits files, runs commands and tests, and reports back. They differ mainly in whose models and plans they use.

Claude Code is Anthropic's agent. Its documentation describes it as "an agentic coding tool that reads your codebase, edits files, runs commands, and integrates with your development tools". It runs in the terminal, in VS Code and Cursor, in JetBrains IDEs through a plugin, in a desktop app and on the web, and you can mention @Claude in Slack with a bug report and get a pull request back.

Codex is OpenAI's agent. It comes as a command-line tool whose GitHub repository is licensed under Apache-2.0, an extension for VS Code and compatible editors such as Cursor and Windsurf, part of the ChatGPT desktop app for macOS, Windows and Linux, and SDKs for TypeScript and Python. On GitHub you can comment @codex review on a pull request. The Codex CLI also has an /import command that brings supported setup and recent chats across from Claude Code or Cursor.

How do Codex and Claude Code compare feature by feature?

Feature for feature, they are now close. Both offer the same building blocks; the differences are in plan access, cloud providers and how each one is configured.

Codex vs Claude Code: what the official docs list (16 November 2026)
FeatureCodexClaude Code
Terminal CLIYes, Apache-2.0Yes
VS Code and CursorYesYes
JetBrains IDEsYes, via AI ChatYes, plugin
Desktop appYes, macOS, Windows, LinuxYes, Linux in beta
Cloud tasksYesYes
GitHub pull request reviewYes, @codex reviewYes, Team and Enterprise
Project instructionsAGENTS.mdCLAUDE.md or AGENTS.md
Skills, hooks, subagentsYesYes
Plugins and MCP serversYesYes
OS-level sandboxYes, on by defaultYes, opt-in
Admin-enforced settingsYes, managed requirementsYes, managed settings
Amazon BedrockYes, local surfacesYes
Google Cloud, Microsoft FoundryNot documentedYes
Free plan accessYes, limitedNo

Sources: OpenAI, Codex documentation, 2026; Anthropic, Claude Code documentation, 2026

Two rows matter most. First, project instructions: Codex reads AGENTS.md files before doing any work, and Claude Code can read a repository's AGENTS.md in place of CLAUDE.md, so one shared file can steer both. Second, cloud providers: Claude Code documents Amazon Bedrock, Google Cloud's Agent Platform and Microsoft Foundry, while Codex documents Amazon Bedrock for its local surfaces. If your company buys AI through one of those clouds, that may decide the question for you.

How much do Codex and Claude Code cost?

Entry prices are similar; the difference is in what each plan's usage allowance buys. At the time of checking (November 2026), OpenAI's Codex pricing page says Codex is included in ChatGPT Free, Go, Plus, Pro, Business, Edu and Enterprise, with Go at $8 a month, Plus at $20 a month and Pro "at $100, $200, or $500 USD per month". Business is $20 per user per month billed annually for two or more users, or $25 billed monthly.

Anthropic's pricing page lists Claude Pro at $20 billed monthly ($17 a month on an annual plan), Max at $100 and Ultra at $200 a month. Team standard seats are $20 per seat billed annually or $25 monthly, and premium seats $100 or $125. Enterprise is listed at US$20 per seat per month, billed annually with a 20-seat minimum, plus usage. The free Claude plan does not include Claude Code. Both vendors publish these prices in US dollars, and Anthropic's page says "Prices shown don't include applicable tax", so budget for VAT.

Published monthly prices for plans that include each agent

US dollars per person or seat, before tax. Low to high across the tiers or billing options shown.

PlanFrom – To
ChatGPT Go to Pro (Codex)$8 – $500
Claude Pro to Ultra$17 – $200
ChatGPT Business seat$20 – $25
Claude Team standard seat$20 – $25
Claude Team premium seat$100 – $125

Sources: OpenAI, Codex pricing, checked November 2026; Anthropic, Claude pricing, checked November 2026

Usage limits are where the real cost sits. OpenAI's page says the number of messages "depends on the model used, size and complexity of your tasks", that local and cloud work share one allowance and that weekly limits may also apply. Anthropic also sets usage by plan, and subagents count toward the same limits as the main conversation. Our guide to Claude Code plans and team setup goes further on seat choice.

Which performs better on coding benchmarks?

On the one public leaderboard that ranks both agents as products, Claude Code entries currently lead Codex entries by about five points, with overlapping error margins. Terminal-Bench 4.1, hosted by Stanford, Harbor and the Laude Institute, reports the share of its tasks each agent and model pair resolves.

Terminal-Bench 4.1 resolution rate, top entries for each agent

Each entry 330 trials at maximum reasoning effort. Margins of about ±3 points.

Agent and model (entry date)Resolved
Claude Code, Opus 5.5 (Sep 2026)65.8%
Claude Code, Fable 5.2 (Nov 2026)65.8%
Codex, GPT-6.1 Sol (Sep 2026)60.9%
Codex, GPT-6 Astra (Sep 2026)60.3%

Source: Terminal-Bench, Terminal-Bench 4.1 leaderboard, checked 16 November 2026

Read these numbers with care. The leaderboard lists Anthropic and OpenAI as the organisations behind these agent and model pairs, and it does not say who ran each entry, so treat them as vendor results rather than independent testing. The margins shown (±3.1 points for the top Claude Code entry, ±3.0 for the top Codex entry) mean the true gap could be small. SWE-bench Verified, the other benchmark often quoted, does not list Codex or Claude Code as agents, and its newest entries dated from February 2026 when we checked. Neither tells you how an agent handles your codebase and tests.

How widely are Codex and Claude Code used?

Claude Code was the fourth most used out-of-the-box agent among developers who use or build AI agents in the Stack Overflow Developer Survey 2025. Codex was not listed as an answer, so the survey cannot compare the two directly.

Out-of-the-box AI agents and assistants used by developers who work with agents

Base: 8,323 respondents who use or develop AI agents.

ToolUsed
ChatGPT82%
GitHub Copilot68%
Google Gemini47%
Claude Code41%
Microsoft Copilot31%

Source: Stack Overflow, Developer Survey 2025: AI

The same survey found trust is still low: more developers said they actively distrust the accuracy of AI tools (46%) than trust it (33%). That is the practical case for keeping tests and human review as the gate, whichever agent you choose.

How do they handle security and permissions?

Both sandbox the commands an agent runs, and both can be configured unsafely. The defaults differ: Codex sandboxes local work by default, while Claude Code's sandbox is something you turn on and its permission prompts do the work otherwise.

Codex's documentation says that by default "the agent runs with network access turned off". In version-controlled folders the default is Auto, meaning workspace write with on-request approvals, enforced on macOS by Seatbelt and on Linux by bwrap plus seccomp. A full-access mode removes the sandbox and approvals, and OpenAI labels it not recommended. Admins can set managed requirements, which the docs describe as "admin-enforced constraints that users can't override".

Claude Code uses permission modes, from asking on first use of each tool to accepting edits automatically, plus allow, ask and deny rules. In its words: "Rules are evaluated in order: deny, then ask, then allow." Its sandbox uses Seatbelt on macOS and bubblewrap on Linux and WSL2. Managed settings sit above every other level, so no user or project setting overrides them, apart from a few documented exceptions.

The real risk in both is an agent with broad permissions on a machine that holds production credentials. Keep secrets out of the repository and require review before anything merges. Our security checklist for AI-built apps covers the checks worth running on agent-written code.

Where are your code and prompts processed, and what does that mean under UK GDPR?

Both agents run on your machine but send prompts, code context and outputs to the vendor's models, so the data terms of your plan apply. For commercial plans, neither vendor trains on your business data by default.

Anthropic's Claude Code data usage page says that "Claude Code runs locally" but sends "all user prompts and model outputs" over the network, encrypted via TLS 1.2 or later. On Team, Enterprise and API accounts, Anthropic "does not train generative models using code or prompts sent to Claude Code under commercial terms", and the standard retention period is 30 days. Consumer Pro, Max and Ultra accounts are different: training depends on a user setting, with 5-year retention if it is on. That alone is a reason to put business users on Team or Enterprise. On the direct Claude API, Anthropic's data residency documentation says only US and global inference regions are available, and data at rest is stored in the US. Anthropic's regional compliance page lists European regions through Amazon Bedrock and Google Cloud, both of which Claude Code supports, so a cloud route is the usual option for teams that need processing outside the US.

OpenAI's Codex enterprise documentation says that for supported Business, Enterprise and Edu workspaces, business data "is not used to train or improve OpenAI models by default" and is encrypted in transit and at rest. It also notes that data residency and inference residency "cover only eligible content and supported workloads, regions, and configurations", so confirm the regions for your workspace with OpenAI before rollout.

Under UK GDPR, sending personal data in code, logs or test fixtures to a provider outside the UK can be a restricted transfer. The ICO's guide to international transfers explains when the transfer rules apply. The simplest control is to keep real personal data out of the repositories agents can read. The NCSC's guidelines for secure AI system development are a useful frame for the rest: secure design, development, deployment, and operation and maintenance.

Which is easier to customise?

In November 2026 both document the same building blocks, so the choice is about fit rather than capability.

In Claude Code, skills are folders with a SKILL.md file that Claude uses when relevant, hooks are "user-defined shell commands" that run at points in its lifecycle, and each subagent "runs in its own context window with a custom system prompt, specific tool access, and independent permissions". Plugins bundle these, and MCP servers connect your systems. Codex now offers hooks (generally available, with events such as PreToolUse and PostToolUse), skills, subagents, plugins and MCP servers configured in config.toml. Our explainer on how MCP connects agents to your business systems covers the integration side.

Claude Code vs Cursor: what is the difference?

Cursor is a code editor with its own agent; Claude Code and Codex are agents that can run inside Cursor through their extensions. If your team wants AI help as they type, Cursor fits. If you want an agent that can be scripted, shared and governed across a team, Claude Code or Codex fits. Our comparison of Lovable, Bolt, Replit and Cursor covers Cursor alongside the app builders.

Which coding agent should your business use?

Pick the agent that fits the subscriptions, cloud and data terms you already have, then confirm it with a trial on your own code. As a starting point:

  • Choose Codex if your company already runs on ChatGPT Business or Enterprise, you want agent work available to people outside engineering in the ChatGPT app, or you want a sandbox that is on by default.
  • Choose Claude Code if you buy AI through Google Cloud or Microsoft Foundry, you want Code Review on Team or Enterprise, or you already use Claude for your wider team.
  • Use both if different teams prefer different tools. Put shared rules in AGENTS.md, keep tests and code review as the gate, and compare results on the same tasks.
  1. Pick three real tasks: a bug fix, a small feature and a refactor
  2. Write one shared AGENTS.md with your conventions and test command
  3. Run each task in both agents on the plan you would buy
  4. Compare review time, test results and usage consumed
  5. Set permissions and managed settings before rolling out
A fair two-week trial of Codex and Claude Code

Whichever you pick, the gains come from the setup around it: clear project instructions, a test suite the agent can run, narrow permissions and a review step. For background, read what agentic AI is and how it works. If you want help, our Claude Code training for teams covers instructions, permissions and review habits, and our AI agent development work goes beyond coding into business workflows.

Circlemount is based in Telford, Shropshire, where Marcus (founder) and a team of 12 specialists set up coding agents for teams across the UK, from Shrewsbury and Wolverhampton to Manchester and London. For SaaS and technology companies, data handling and which cloud the agent runs through are worth settling before any tool is installed.

Frequently asked questions

Is Codex better than Claude Code?

Neither is better for every team. On the Terminal-Bench 4.1 leaderboard the top Claude Code entry scored 65.8 percent and the top Codex entry 60.9 percent, with margins of about three points. Plan access, cloud provider and data terms usually matter more, so trial both on the same real tasks before standardising.

Is Codex free to use?

As of November 2026, OpenAI says Codex is included in ChatGPT Free, Go, Plus, Pro, Business, Edu and Enterprise. The Free plan is aimed at quick coding tasks, and usage limits depend on the model and the size of your tasks, so regular daily use usually needs a paid plan. Check the current pricing page before buying.

Do I need a paid plan for Claude Code?

Yes. Claude Code needs a Pro, Max, Ultra, Team or Enterprise plan, or an Anthropic Console account billed by usage; the free Claude plan does not include it. Organisations can also run it through Amazon Bedrock, Google Cloud's Agent Platform or Microsoft Foundry using their existing cloud agreements and billing.

Can Codex and Claude Code share the same project instructions?

Yes. Codex reads AGENTS.md files before doing any work, and Claude Code can read a repository's AGENTS.md in place of CLAUDE.md. Many teams keep shared rules for structure, testing and style in AGENTS.md and add tool-specific skills, hooks and settings separately, so both agents follow the same conventions.

Do Codex and Claude Code train on my company's code?

Not by default on business plans. Anthropic says it does not train on code or prompts sent to Claude Code under commercial terms, such as Team, Enterprise and API. OpenAI says business data in supported Business, Enterprise and Edu workspaces is not used for training by default. Consumer plans have different settings, so check them.

Is it safe to let a coding agent run commands?

It is reasonably safe inside a sandbox with narrow permissions. Codex sandboxes local work by default, and Claude Code offers an operating-system sandbox plus allow, ask and deny rules. Run agents in a development environment, keep production credentials out of reach, block risky commands and review every change before it merges.

Should I use Cursor or Claude Code?

You can use both. Cursor is a code editor with its own agent, and Claude Code and Codex can both run inside it through their extensions. Cursor suits developers who want help as they type. Claude Code suits teams that want a scriptable, shareable agent with central rules and permissions.

Sources

Checked on 16 November 2026. Product features and figures change quickly, so check the current source before you rely on a detail.

Want help rolling out a coding agent?

As a Claude Code agency, Circlemount sets up Claude Code and Codex for UK teams: project instructions, permissions, skills, hooks, data handling and the review process that keeps quality up. Join the Circle for a free business analysis and proposal.